Lightweight Architectures for Binary Segmentation of Fresh Mud Deposits in UAV Imagery

Abstract

Monitoring surface activity in mud volcanoes is critical for understanding fluid-driven geological processes and mitigating associated hazards. This work presents a deep learning-based approach to segmenting nadiral imagery captured by Unmanned Aerial Vehicles (UAVs), with the aim of identifying fresh mud deposits. Given the scarcity of labeled data and the absence of pretrained models suitable for this domain, we developed custom segmentation architectures tailored to the visual complexity and scale of UAVcaptured volcanic terrain. The proposed models were trained from scratch and evaluated on a dataset collected across active sites in Sicily. Results show that domain-adapted network design enables robust delineation of geologically relevant features, offering a practical framework for automating UAV-based observation in environmental and geohazard monitoring.